Total daily streamflow in yield (mm), over the period of record
5.2 Sensitivity analyses
5.2.1 Missing data
Many of the EcoDrought time series data are incomplete. At some sites, discharge data is available only during the summer and/or fall periods, and at other sites, time series data are interrupted due to malfunctioning sensors and/or ice formation (“ice spikes”). So how does the length of the time series affect baseflow separation (and subsequent event identification)? Wasko and Guo (2022) use a 67 day time series of flow to demonstrate the utility of the hydroEvents packages, suggesting digital baseflow separation techniques may be valid for relatively short time series.
Here, I perform a simple sensitivity analysis to explore the effect of time series length on the results of baseflow separation. Essentially, perform baseflow separation on increasingly smaller subsets of the data. With the default parameters, the minimum number of days/observations needed is 31. This is because the default number of points reflected at start and end of data (r) is 30. Reflection allows bf/bfi to be calculated over the entire period of record as the underlying baseflow separation equations result in “issues of”Warm-up” and “cool-down” as the recursive filter is moved forward and backward over the dataset” (Ladson et al. 2013, Australian Journal of Water Resources). baseflowB() uses a default reflection period of 30, which Ladson et al. (2013) found to “provide a realistic baselfow response for the start and end of the actual flow data”.
Divergence in baseflow among datasets is a result of the reflected data of the shorter dataset not matching the actual data of the longer dataset. As a result, divergence really only occurs at the end of each time series and is generally small in magnitude.
Pairs plot of baseflow derrived from datasets of different lengths. Red lines are 1:1.
5.2.1.2 Compare baseflow index
The story here is essentially the same as above: divergence is ~minimal and restricted to the end of each time series. However, we note that divergence in BFI appears to increase as absolute flow/baseflow decreases, because small differences in absolute space become much larger in relative space when absolute values are small.
Pairs plot of baseflow derrived from datasets of different lengths. Red lines are 1:1.
5.2.2 Time scale
Given that streamflow can change so quickly in small, headwater streams, are we missing a key part of the story by using flow data summarized as daily means? Using daily mean flow reduces the range of values, particularly at the upper end (i.e., high flows), and so we may be overlooking the g~G relationship at very high flows.
5.2.2.1 Organize data
First, set baseflow separation (for the Lyne-Hollick one-parameter digital recursive filter) and event delineation paramters (as in Wasko and Guo, 2022)
alp: alpha filter parameter, higher values “lower” the estimated baseflow (thus making it difficult to delineate events)
numpass: number of passes. Ladson et al. (2013) recommend 3 passes for daily data (default in baseflowB() function) and 9 passes for hourly data
thresh: baseflow index threshold for event delineation, higher threshold values make it “easier” to delineate events
Code
alp <-0.925numpass <-9thresh <-0.75
Download 15-min NWIS data for big G (West Brook NWIS)